Asynchronous Federated Caching Strategy for Multi-Satellite Collaboration Based on Deep Reinforcement Learning

By incorporating caching functions into Low Earth Orbit (LEO) satellites, users worldwide can benefit from caching services. However, satellite caching faces the following challenges: 1) The continuous mobility of satellites introduces dynamic shifts in user distribution, resulting in unpredictable variations in interested content over time. 2) The cached content is susceptible to becoming obsolete due to the brief connection times established between satellites and clients. 3) Significant concerns arise regarding data privacy and security. Users may exhibit reluctance to transmit local data for privacy protection. To address the abovementioned challenges, we propose an asynchronous federated caching strategy (AFCS) consisting of an access satellite and several collaboration satellites. Clients employ an asynchronous federated learning methodology to collaboratively train a global model for predicting content popularity. Concerning privacy protection, clients are not required to upload local data. Instead, they only need to transmit the model hyperparameters. This approach significantly diminishes the risk of data leakage, thereby safeguarding data privacy effectively. We propose a novel strategy for client selection participating in global model training. Through model training, we can get a preliminary caching strategy. To further improve caching performance, we propose a multiple-satellites collaboration based on deep reinforcement learning. This collaborative approach enhances the cache hit ratio and diminishes content request delay.